Chapter 6: Multiple Regression I

6.1 Multiple regression models

3-D Algebra Review

Plot Y = 1 + 2x1 - 3x2

Y / X1 / X2
1 / 0 / 0
3 / 1 / 0
-2 / 0 / 1
0 / 1 / 1

The connected (Y, X1, X2) triples form a plane:

From above:

From the X1 and X2 plane level (at Y=-3):

Simple linear regression

One predictor variable is used to estimate one dependent variable.

Multiple linear regression

Two or more predictor variables are used to estimate 1 dependent variable.


1)i~independent N(0,2)

2)0, 1, …,p-1 are parameters with corresponding estimates of b0, b1, …, bp-1

3)Xi1,…,Xi,p-1 are known constants (see Section 2.11 when the X’s are random variables – mostly everything stays the same)

4)The second subscript on Xik denotes the kthpredictor variable.


Now suppose there are only 2 predictor variables:


1)Often when we want to just refer to the first, second,… predictor variables, the i subscript is dropped from Xij. Thus, predictor variable #1 is X1, predictor variable #2 is X2,…

2)Yi=0+1Xi1+…+p-1Xi,p-1+i is called a “first-order” model since the model is linear in the predictor (independent, explanatory) variables.

3)There are p – 1 predictor variables and p  parameters.

4)The coefficient on the kthpredictor variable is k. This measures the effect Xk has on Y with the remaining variables in the model held constant.

5)Qualitative variables (variables not measured on a numerical scale) can be used in the multiple regression model. For example, let Xik=1 to denote female, Xik=0 to denote male. More will be done with qualitative variables in Chapter 8.

6)Polynomial regression models contain higher than first order terms in the model. For example, . More will be done with polynomial regression models in Chapter 7.

7)Transformed variables can be used just as in simple linear regression. For example, log(Yi) can be taken to be the response variable. More will be done with transformed variables in Chapter 7.

8)If the effect of one predictor variable on the response variable depends on another predictor variable, interactions between predictor variables can be included in the multiple regression model. For example, . More will be done with interactions in Chapter 7.

9)The word “linear” in multiple linear regression is in reference to the parameters. For example, is not a multiple LINEAR regression model.

General linear regression model in matrix terms

Note that the Yi=0+1Xi1+…+p-1Xi,p-1+i for i=1,…,n can be written as:






Then Y = X + , which is

Remember that  has mean 0 and covariance matrix

since the i are independent.

Note that E(Y) = E(X + ) = E(X) = X since E() = 0

Question: What is Cov(Y)?

Estimation of regression coefficients (j’s)

Parameter estimates are found using the least squares method.

From Chapter 1: The least squares method tries to find the b0 and b1 such that SSE = (Y-)2 = (residual)2 is minimized. Formulas for b0 and b1 are derived using calculus.

For multiple regression, the least squares method minimizes SSE = (Y-)2 again; however, is now

As shown in Section 5.10, the least squares estimators are . This holds true for multiple regression with

Properties of the least squares estimators in simple linear regression (such as: unbiased estimators and minimum variance among unbiased estimators) hold true for multiple linear regression.

6.4 Fitted values and residuals

Let . Then since


Hat matrix: where H=X(XX)-1X is the hat matrix

We will use this in Chapter 10 to measure the influence of observations on the estimated regression line.

Covariance matrix of the residuals:

Cov(e) = 2(I-H) and

 2012 Christopher R. Bilder